It may generally be less likely to encounter disk read errors than it is to encounter network timeouts, but both will eventually occur.
The abstraction works well. The main issue is when you lose access to your root device. Just like if you booted a Linux machine from NFS, the root disk connection must be stable (how do you unmount without being able to read the binary). All other mounts can be arbitrarily flakey.
We have enough RAM now that there's no reason (even on embedded devices!) to ever unload the initrd/initramfs. An OS should be able to be configured such that you can unmount your rootfs whenever you like, and just automatically be de-pivot-root'ed and end up back in your initramfs, where you can mount the rootfs again.
Anyway, you could make a ramdisk, put things in it and put it in your path. I doubt that a linux box will survive a dead root without hardcore sysadm skills, but it should work on plan9. Initrd is not meant to stay behind after boot, though.
Does this kind of abstraction actually work well? Networks are generally quite unreliable so does it really work well to treat everything as files even when the latency can be so different depending on if the resource is local or remote? Do the plan9 APIs make it natural to write code that deals well with random blocking when reading/writing to files?